Jun 15, 2026 · Our paper What Do Evolutionary Coding Agents Evolve? introduces EvoTrace, a dataset of 121 evolutionary code-search runs, and EvoReplay, a tool for inspecting and rerunning parts of those runs. The goal is to tell whether better scores come from new program structure, constant tuning, context changes, overfitting, or repeated code churn.
Apr 12, 2026 · Our paper The Agentic Researcher presents an open-source framework that turns CLI coding agents into autonomous research assistants. Through methodological rules formulated as agent prompts, we enable sustained autonomous research sessions — including sessions running over 20 hours across multi-node GPU clusters without human intervention.